Investigation of acoustic features related to speech intelligibility and its application as an evaluation method of noise reduction algorithms

Abstract

To understand the sequence of speech intelligibility, researchers specializing in phonetics and acoustics are interested in understanding the relationship between acoustic features and speech intelligibility. These studies were performed under the assumption that specific acoustic features play an important role in speech intelligibility, and if these specific acoustic features can be identified, it can be applied to improve speech intelligibility. The aforementioned studies have demonstrated that specific acoustic features such as fundamental frequency, formant frequency, spectral energy, amplitude modulation, and envelope are highly correlated with speech intelligibility. Previous studies have only used single acoustic features in their experiments and it is impossible to compare or rank these single acoustic features mentioned in those previous studies as they are each performed in different environments. Hence, a study with multiple acoustic features with identical conditions is needed to compare and rank the usefulness of the acoustic features in speech intelligibility. In addition, previous studies were performed in English and Mandarin. But there is no related study based in Korean which has different phonetics and acoustics compared to English and Mandarin. This thesis aims to investigate acoustic features related to speech intelligibility in Korean and English. Multiple correlation analysis is performed on acoustic features, thought to be related to speech intelligibility by previous studies, to identify acoustic features with high correlation to speech intelligibility. And identify acoustic features were applied on performance evaluation of nine noise reductions that are used commercial filed for analyzed quantitatively to relation with selected acoustic features and speech intelligibility after noise reduction process. The Result showed that four acoustic features - AM(w) (Amplitude modulation of 4~16Hz for word), SB(w) (spectral balance for word), FRT12(w) (Formant frequency ratio F1/F2 for word), FRT12 (s) (Formant frequency ratio F1/F2 for sentence) - are highly correlated (r >0.60) to speech intelligibility in a variety of situations. And the order of correlation from high to low were AM(w) (r >0.75), FRT12(w) (r >0.67), FRT12(s) (r >0.65), and SB(w) (r >0.64). Of the four acoustic features, AM(w) and SB(w) were found to be proportional to speech intelligibility of the noise reduction processed output. After statistical analysis, this result was more apparent in Korean than in English. This thesis have demonstrated that the identified four acoustic features have an important role in speech intelligibility. These results can be applied to the development of speech processing algorithm and speech recognition systems. Also, AM(w) and SB(w) features can replace the subjective speech intelligibility measurement to evaluate the performance of speech signal processing algorithms.